4DR P2T: 4D Radar Tensor Synthesis with Point Clouds

TL;DR

The paper tackles the limitation of CFAR in 4D Radar by proposing 4DR P2T, a cGAN-based framework that synthesizes tensor data from 4D Radar point clouds to preserve spatial object characteristics for deep learning. It introduces a 3D encoder–decoder architecture with a 3D multi-scale discriminator and a composite loss combining L1, perceptual, and conditional adversarial terms, evaluated on the K-Radar dataset. The approach achieves a strong average PSNR of $30.39$ dB and SSIM of $0.96$, with percentile-based data reductions identifying a 5% percentile as best for tensor fidelity and 1% percentile as optimal for data-volume efficiency, guiding DL training. This work enables high-fidelity tensor representations for radar data, supporting improved perception and sensor fusion in autonomous driving, with future plans to include Doppler and unpaired-data extensions.

Abstract

In four-dimensional (4D) Radar-based point cloud generation, clutter removal is commonly performed using the constant false alarm rate (CFAR) algorithm. However, CFAR may not fully capture the spatial characteristics of objects. To address limitation, this paper proposes the 4D Radar Point-to-Tensor (4DR P2T) model, which generates tensor data suitable for deep learning applications while minimizing measurement loss. Our method employs a conditional generative adversarial network (cGAN), modified to effectively process 4D Radar point cloud data and generate tensor data. Experimental results on the K-Radar dataset validate the effectiveness of the 4DR P2T model, achieving an average PSNR of 30.39dB and SSIM of 0.96. Additionally, our analysis of different point cloud generation methods highlights that the 5% percentile method provides the best overall performance, while the 1% percentile method optimally balances data volume reduction and performance, making it well-suited for deep learning applications.

Figures (4)

• Figure 1: 4DR P2T overview. The 4DR P2T model generates tensor data from 4D Radar point clouds, which are represented in bird’s-eye view (BEV) as a 2D projection. Traditional point cloud generation methods often suffer from measurement loss, which may affect their suitability for deep learning training. To mitigate this limitation, the model generates tensor data to prevent measurement loss, ensuring that crucial information is retained for deep learning tasks.
• Figure 2: 4D Radar signal processing and data representation 4D_radar_survey4dradar_tutorial4dradar_data_representation. The Radar power values are normalized and represented using colors. The Radar point cloud is shown as black points, and the bounding box for the objects is indicated with a red box.
• Figure 3: Overall structure of 4DR P2T. The encoder utilizes 3D sparse convolution to process 4D Radar point cloud data, while the decoder employs 3D dense convolution to generate tensor data.
• Figure 4: Qualitative experimental results of 4DR P2T. The top part shows the front camera image and LiDAR point cloud as reference data to understand the scene of the 4D Radar GT tensor data, while the bottom part presents the tensor data results generated by 4DR P2T under different point cloud generation methods and density conditions.